Lying mirror

We present an all-optical system, which we term “lying mirror”, designed to conceal visual input information by transforming it into standardized, ordinary-looking output patterns. This optical transformation is executed passively through the interaction of incident light with an optimized, structured diffractive surface, enabling optical obfuscation without digital computation. Designed using deep learning, the lying mirror spatially modulates incident wavefronts to project a consistent, “dummy” output image regardless of the original input content, effectively camouflaging sensitive visual data. The designs are shown to hide various classes of input data while exhibiting robustness to adversarial manipulations such as random noise, rotations, translations, and scaling of object features. We validated the system’s feasibility experimentally with a structured MEMS-based micro-mirror array illuminated by multi-wavelength light covering the blue, green, and red channels. Additionally, we engineered a broadband lying mirror capable of operating over a continuous spectral range, increasing its adaptability to different illumination sources.